Joint training of retrieval representations lifts all boats, BGE-m3 and NeoMME findings
antoine_chaffin · x · 2026-10-14
Citing BGE-m3 (2024) and NeoMME (2026), the author argues that jointly training dense/sparse/multi-vec retrieval representations outperforms parallel/sequential training. Attached figures show self-distillation ablation gains (sparse 36.7→53.9 nDCG@10) and NeoMME-260M dual-head joint training beating single-head on ViDoRe v3 and BEIR-15. Call for more joint training of late-interaction models.
More from Research
- Submarine cyber ops study: operator cognitive capacity swings with the tactical environment — J_MaestreVidal · 2026-10-14
- Recova Trains Robots to Self-Recover in Simulation, Cutting Human Intervention From 87.5% to 0% — Lianhuiq · 2026-10-14
- NVIDIA's AlphaNIST Learns Robotic Assembly From Scratch With Only Binary Rewards, Hitting 98.2% Across 20 Tasks — abhishekunique7 · 2026-10-14
- Encrypted Chain-of-Thought Can Be Decoded: Reasoning Theft Attack on Top LLM APIs — matthew_d_green · 2026-10-14
- Most people carry their mother's cells in their brain, even at 80 — AnnaCiaunica · 2026-10-14
- Looping Layers Boosts Small Decision Models: 1.4B Nears 3x Larger Model — rohanpaul_ai · 2026-10-14